EPISTEMIC SKILLS (2026)

An empirical study and tooling suite for authoring SKILL.md specifications that produce measurable causal lift on agent benchmarks rather than context bloat. Across a 13-task stratified sweep on the BenchFlow harness (n = 1000 paired bootstrap), the 11-skill static library raised resolution from 15.4% (no-skill baseline) to 30.8% (+15.4pp).

The repository codifies six empirical authoring laws derived from trajectory autopsies (such as enforcing compact procedural checklists under 600 words and embedding executable verification gates) into a deterministic linter and packager. For self-generated skills, a standalone claim verifier parses quantitative and structural claims out of candidate outputs and blocks task completion until every claim passes programmatic inspection.